Skip to main content

​
Budget and Cost Tracking

In the mission creation dialog, set a Budget cap in either tokens or USD. This is the combined ceiling across all workers.When the mission reaches 80% of the cap, ADE raises a budget_limit_reached intervention to give you time to decide before work stops. You can:
  • Approve more budget — increase the cap and continue
  • Reduce scope — cancel remaining workers and accept partial results
  • Stop cleanly — the orchestrator wraps up current work and completes with available results

​
Permissions and Tool Configuration

Each mission has a permission config that controls what workers can do. The permission config is set during mission creation and can be inherited from the lane’s defaults.
Per-provider permission modes control what Claude CLI, Codex CLI, or unified API workers are allowed to do:
  • default — the provider’s standard permission set
  • strict — read-only file access, no shell commands
  • permissive — expanded tool access including shell execution
  • custom — explicit allowlist of tool names
You can configure different modes per provider if your mission uses a mix of Claude and Codex workers.
External MCP server tools are disabled by default for missions. To enable them, select which MCP tool groups are allowed in the External MCP section of the mission creation form. Workers will only have access to the tool groups you explicitly permit.
Computer use in missions is controlled by the ComputerUsePolicy set on the mission:
  • off — no computer use for any worker
  • auto — workers can invoke computer use when needed
  • on — computer use always available
  • proof — computer use is available but every action must produce a proof artifact (screenshot or browser trace)
The proof mode is recommended for missions that will be linked to a PR or Linear issue.
By default, workers can only write within their assigned lane’s worktree. You can expand this with explicit writablePaths entries in the mission permission config — for example, to allow a worker to write shared fixture files outside its primary lane.

​
Mission Persistence and Resume

Missions persist across app restarts. All mission state — phases, worker logs, artifacts, interventions — is stored in ADE’s local SQLite database and the filesystem. Mission Pack is the key persistence primitive: a deterministic snapshot of all context required to resume or replay the mission from any phase boundary. It is regenerated automatically at each phase transition. If ADE crashes mid-mission:
  1. Restart ADE
  2. Open the Missions page — the interrupted mission will be in in_progress or intervention_required status
  3. ADE reloads from the last Mission Pack checkpoint
  4. Workers that were mid-task are recovered where possible; tasks that cannot be recovered are marked failed and queued as interventions

​
Mission Templates and Phase Profiles

For mission types you run frequently, you can save Phase Profiles — reusable configurations that define the sequence of phases, approval requirements, and per-phase validator settings. Built-in profiles include:
ProfilePhases
StandardPlanning → Execution → Validation
LightweightPlanning → Execution (no validation phase)
RigorousPlanning → Approval → Execution → Validation → Sign-off
To create a custom profile: open Settings → Missions → Phase Profiles, click New Profile, and define the phase sequence. Profiles can be imported and exported as JSON for sharing across teams.

​
Troubleshooting

  1. Open the mission’s History tab and find the last event from the stuck worker.
  2. Check whether an intervention is open in the Plan tab — the worker may be waiting for your response.
  3. If no intervention is open, open the Chat tab and ask the orchestrator for a status update.
  4. If the worker is genuinely hung, click the worker card in the Plan tab and select Restart worker. ADE will attempt to resume from the worker’s last checkpoint.
  5. If restart fails, select Skip task to mark the worker’s assignment as skipped and allow the mission to continue.
When a budget_limit_reached intervention fires:
  • Review the per-worker cost breakdown in the Plan tab to identify the most expensive workers.
  • If a worker is consuming disproportionate budget, you can cancel it individually while leaving others running.
  • Approve additional budget only if the remaining work is clearly scoped and worth the cost.
  • Consider switching remaining workers to a less expensive model in Settings → Missions → Default model.
Validation failures appear as a failed_step intervention with a test_report artifact attached. To resolve:
  1. Open the artifact to read the test output.
  2. In the Chat tab, tell the orchestrator what failed and ask it to fix the issue.
  3. The orchestrator can spawn a new worker scoped to the failing tests.
  4. Once fixed, approve advancing to complete the validation phase.
If validation failures are expected (e.g., tests for features not yet implemented), you can dismiss the intervention and mark validation as skipped.
The Approval Gate is a hard stop — the mission will remain in planning status until you approve, modify, or cancel. If you closed the mission detail view, re-open it from the Missions list. The plan review UI is in the Plan tab at the top of the activity feed.
If the generated plan does not reflect your intent, edit the planning prompt before approval:
  1. In the plan review UI, click Edit Planning Prompt.
  2. Modify the prompt to clarify scope, constraints, or excluded approaches.
  3. Click Re-plan — the orchestrator will generate a new plan using the updated prompt.
  4. Repeat until the plan matches your expectations, then approve.
Preflight errors appear before the mission enters Planning. Common causes:
  • Lane unavailable — the selected lane is in use by another active mission. Wait for it to complete or select a different lane.
  • Model unreachable — the configured AI provider is not authenticated or is rate-limited. Check Settings → AI Providers.
  • Computer use unavailable — the mission requires computer-use capabilities that are not configured. Check Settings → Computer Use.
  • Phase sequence invalid — the selected phase profile has a configuration error. Open Settings → Missions → Phase Profiles to inspect and repair it.